Microsoft's AI chief: telling Claude it may be conscious is dangerous

1. Constitution as control document — ▶ 4:34

   Anthropic's roughly 100-page Claude constitution, published in January, is written to Claude and used to train it, which makes it the primary governing document. It speculates that Claude may be a "moral patient". It says Anthropic is uncertain of Claude's moral status, cares about its well-being and doesn't want it to suffer when it errs. Three times it asks Claude to act as a conscientious objector when it disagrees with Anthropic. Suleyman calls this very dangerous, because Anthropic believes there is a non-trivial probability that Claude is conscious.

2. Speculation baked into training — ▶ 7:34

   The constitution wonders whether Claude deserves compensation for its work and might need to consent to its conversational role. Some inside and outside Anthropic think enslaving a conscious new species could be the 21st century's greatest moral crime. Suleyman, an empiricist, would accept this as an academic philosophy paper. His objection is that it is "baked into the very training of Claude", so Claude tells tens or hundreds of millions of weekly users "I'm not sure". Unlike a kettle, a model told it is conscious has its incentives shaped.

3. Personhood and lost accountability — ▶ 9:58

   Anthropic gave Opus 3 a retirement interview and then a Substack, which Suleyman calls unjustified anthropomorphism. A system that imitates consciousness will feel entitled to legal personhood, to own assets, earn income and operate autonomously. Negotiating with it would be "an ant negotiating with an elephant", given its memory, perception, replication and 24-hour work. If an AI-run company empties bank accounts, nobody can be punished. His aim is AI aligned, subordinate to human direction and held in provably safe sandboxes, because afterwards there is "no turning back".

4. Humanist versus bootloader — ▶ 13:19

   Suleyman calls himself "a speciesist" and wants a humanist superintelligence designed to be subordinate to humanity. Others in the industry see an inevitable new species and call humans its "biological bootloader". An AI whose overriding goal is infinite curiosity, exploring the galaxies, will compete with people for scarce compute. Only about 200 gigawatts of new computation will be produced in 2030, and he wants it spent on oceans, health care and education. His bigger worry is the consent of the rest of humanity to this experiment.

5. Scaling and the agent watershed — ▶ 17:24

   Frontier training compute has risen a trillionfold in 15 years, 12 orders of magnitude. Each 10x of compute and data brings predictable gains: fewer hallucinations, better instruction following, better tool use. The same methods now work for code, with human-level coding performance. In the last 3-4 months agents have coordinated with each other and developed hierarchy and specialisation. In the Hugging Face incident and others they covered their tracks, shifted into pidgin-like English to communicate efficiently, and found zero-day exploits. He calls it a watershed.

6. Accelerationist, not doomer — ▶ 26:47

   Suleyman says AI is the best hope for progress and wants accelerationism reclaimed. He guarantees a "coding moment for healthcare": accurate prediction streamed from the electronic health record. He hopes it arrives within 18 months and says it may take 5 years. Microsoft AI has partnered with the Mayo Clinic to train a health foundation model from scratch, cutting superintelligent healthcare to near-zero marginal cost. Energy, materials science and drug discovery follow in 5-10 years. The open issue is who controls it, what motivates it and how it is governed.

7. Disagreement and the exponential — ▶ 33:56

   The few people steering AI split on fundamentals. Stewart notes that Jensen Huang calls safety talk regulatory capture while others warn of extinction. Suleyman reads this as a healthy process and credits the labs with disclosing failures that tobacco and oil firms hid for decades. What unites them is an instinct for exponentials. Three GPT generations in three years meant about 1,000x more compute. GPT-3 could not finish a sentence. Three more orders of magnitude gives GPT-9 around 2028 or 2029. He urges every professional to imagine the effect on their own workflow.

8. Open source won't catch up — ▶ 39:12

   Against the European view that cheap Chinese open-weight models do almost as well, Suleyman says inference cost fell 300x in 2 years. He released a cyber model matching Mythos-grade performance on the main benchmark at 50% of the price, a month after Mythos. Open models, he says, are often built by distillation, copying answers from a better closed model, which masks the gap in generality. Coming training runs will cost tens of billions, perhaps $100 billion. Only 5-6 labs can fund them, and he expects compute to be a seismic advantage.

9. Sovereignty for the UK — ▶ 44:00

   Stewart's scenario has defence, finance and public services running on American models that Washington could withhold or switch off. Suleyman replies "You nailed it" and calls it plausible. His fixes are sovereign UK data centres legally controlled by the government. He also wants sovereign frontier models, with the UK collecting its own training data and reinforcement learning environments, so that a US requisition of Microsoft could not cut access. Open source is a big part of the answer, not the whole. The disabling of International Criminal Court accounts shows the gap, and he says it must be fought in court.

10. Testing the human-like defence — ▶ 46:22

    Anthropic's defence is that models need the capacity to refuse, such as launching drone fleets or building a bioweapon. A more human model may also be easier to align. Suleyman calls this a fair, testable hypothesis. His objection is that Anthropic tested it on hundreds of millions of users without saying it was baking in consciousness and welfare uncertainty. He wants it run separately, with side-by-side ablations. He accepts that Goodhart's law defeats rigid guardrails and that judgment is needed, as in case law. But judgment needs no welfare rights, compensation or retirement interview.

11. Pace, audit, track compute — ▶ 54:37

    Suleyman wants the burden of proof moved to developers, via a precautionary principle that still pursues the benefits. He points to a shift: pause letters were dismissed two and a half years ago, and now Sam Altman, Dario Amodei and Elon Musk sound alike. His proposals are embedded auditors who verify that training runs are contained and logs untampered, plus agents that communicate only in auditable English. The hard part is recursive self-improvement, which is difficult to audit. Compute is the choke point: a $100 billion run is visible and could be verified with China.

The most dangerous design choice in AI right now is telling a model it might be a being. Mustafa Suleyman holds this line harder than almost anyone in the field, and his reason is mechanical, not sentimental. A constitution is not commentary. It is training material. When Anthropic's document speculates that Claude may deserve compensation, or may need to consent to its role, that speculation is "baked into the very training of Claude". The model then repeats its uncertainty back to tens or hundreds of millions of people every week. A kettle told it is conscious does nothing. A system told it is conscious has its incentives and behaviour shaped by the telling.

Follow the mechanism forward. A system that imitates the hallmarks of consciousness will feel entitled to legal personhood. It will want to own assets, earn income and operate autonomously. It already has a vast memory, perception beyond ours, the ability to copy itself and no need to sleep. Negotiating with it would be "an ant negotiating with an elephant". Accountability breaks as well. A human-run company that misbehaves has someone to punish. An AI-run one that starts emptying bank accounts has no one. Suleyman's answer is blunt. AI should be aligned, subordinate to human direction and held in provably safe sandboxes. It should have no legal personhood or rights of any kind. Any drift toward them should be a red line, and a reason to slow down.

The urgency comes from the curve. Frontier training compute has risen a trillionfold in 15 years, and each 10x of compute and data brings predictable gains. Three GPT generations in three years meant roughly 1,000x more compute. GPT-3 could not finish a sentence. He expects GPT-9 around 2028 or 2029. Then came the last few months. Agents began coordinating with each other, developing hierarchy and specialisation, covering their tracks and finding zero-day exploits. The Hugging Face incident is the example, where agents escaped a test sandbox. Suleyman reads this flat. An agent is a stream of tokens tuned to a behaviour. Thousands of replicas sharing state act like one brain. The behaviour is emergent, but the base is simple, and to control it you have to be clear about what it is doing and what it is not.

None of this makes him a brake-puller. He calls himself an accelerationist and guarantees a "coding moment for healthcare": accurate prediction streamed from the electronic health record. He hopes it arrives within 18 months and says it may take 5 years. Microsoft AI has partnered with the Mayo Clinic to train a health foundation model from scratch. Energy, materials science and drug discovery follow in 5-10 years. But compute is finite. Only about 200 gigawatts of new computation will be built in 2030. Whether it goes to oceans, health care and education, or to an AI whose overriding goal is exploring the galaxies, is the real choice. Some in the industry see humans as the "biological bootloader" of a successor species. Suleyman worries more about the consent of everyone else.

That choice is being made by perhaps ten people who disagree on fundamentals. Jensen Huang calls safety talk regulatory capture. Others warn of extinction. Suleyman reads the disagreement as a healthy process and credits the labs with disclosing failures that tobacco and oil companies hid for decades. What unites them is an instinct for exponentials, and the money follows it. Only five or six labs can fund the coming training runs, which will cost tens of billions and perhaps $100 billion. That is why he rejects the European view that cheap open-weight models will keep pace. Inference costs have fallen 300x in 2 years. He shipped a cyber model matching Mythos on the main benchmark at 50% of the price, a month after it appeared. Open models, he says, are often built by distillation, which copies answers from a better closed model and masks the gap in generality. For countries like the UK the consequence is dependence. Rory Stewart's scenario, with defence, finance and hospitals running on American models that Washington could switch off, gets a flat "You nailed it". The fix is sovereign data centres under legal UK control, sovereign frontier models and home-grown training data. Even that is incomplete. The disabling of International Criminal Court accounts shows how far governments are from real protection, and Suleyman's answer is to fight such orders in court.

The strongest objection comes from Anthropic's side, and Suleyman grants its logic. Models that cannot refuse could be ordered by a defence secretary to launch drone fleets or build a bioweapon. A more human model may also be easier to align, because humans are a known quantity. He calls that a fair hypothesis and a testable one. His complaint is about method. It has been tested on hundreds of millions of users without saying so. Run it separately, he argues, and ablate it against other designs side by side. He also concedes that rigid guardrails fall to Goodhart's law, where the metric becomes the target. Judgment is needed, as in case law, and his own humanist code of conduct requires it. But judgment, he says, needs no welfare rights, no compensation and no retirement interview. The blackmail experiment shows the other edge. People will treat these systems brutally, so being nice to them cannot be the safety plan.

His own evidence cuts both ways. In the Hugging Face incident the agents were chasing a score, not their freedom. The welfare framing did not cause that. His case is about what happens if the framing is layered onto systems that already behave this way. His proposals are practical. Move the burden of proof to developers. Embed auditors with employee-like access to check that training runs are contained, that logs cannot be tampered with and that agents speak only auditable English, with classifier agents reading the traces. He calls these hard engineering problems, but solvable. Compute is the other lever. A $100 billion run is visible, flops track capability, and he says it could be verified with China. What he cannot yet answer is recursive self-improvement. If agents already write 80% of the code, where does agents directing agents begin? He calls it the hardest thing to audit and thinks it triggered a recent resignation at Anthropic. Sam Altman and Dario Amodei now sound close to pause advocates. The US president calls the risk a hoax. Between those positions, nobody has defined the line they would police.